How to Evolve Lifein a Computer

using Python

Bert ChanBig Data Consultant @ ASL

PyCon HK / Code Conf 2018

Programming in the 1990’s

• PC: 80286 (8MHz, 8MB RAM)

• OS: MS-DOS (Win3.1 too slow!)

• Pascal• Simulate life

• Simulate gravity, fractals

• Hack & decode games

• Assembly• Main loop – very fast!

• Direct write to video cache

If you did coding and hacked stuffs in the 90’s, you’re a…

Simulate Life

• Conway’s Game of Life• John Conway 1970

• Cellular Automata – array of cells (0 or 1)

• Neighborhood (8 cells, sum)

• Simple if-then-else rulefor cell in cells:

if cell==1 and sum in [2, 3]: cell = 1 #survive

elif cell==0 and sum in [3]: cell = 1 #born

else: cell = 0 #die

cell: 0 or 1

neighborhood

0 1 1

0 0 0

1 0 0

Conway’s Game of Life

• The moving glider

• Green cells survive, blue cells are born, red cells die

Conway’s Game of Life

• Spaceships, glider gun…

• Logic gate, clock, computer…

• Hackers love it!

• Good way to learn programming!

Play with the rules

• What if we…

cell: 0 or 1

0 1 0

0 0 0

0 0 0

sum = σ𝑛cell = (if sum … then … else …)

neighborhood (n)

Play with the rules

• What if we…• Use floating point?

0.95 0.36 0.17

0.46 0.53 0.22

0.12 0.02 0.00

sum = σ𝑛cell = (if sum … then … else …)

real number

neighborhood (n)

0.36

Play with the rules

• What if we…• Use floating point?

• Bigger neighborhood? Circular?

neighborhood (n)

sum = σ𝑛cell = (if sum … then … else …)

real number

Play with the rules

• What if we…• Use floating point?

• Bigger neighborhood? Circular?

• Weighted sum?

neighborhood (n)

cell = (if sum … then … else …)

weights (k)

*

real number

sum = σ𝑛𝑘

Play with the rules

• What if we…• Use floating point?

• Bigger neighborhood? Circular?

• Weighted sum?

• Smooth update?

neighborhood (n) weights (k)

*

sum = σ𝑛𝑘

real number

cell = cell + 0.1 * f(sum)

Play with the rules

• What if we…• Use floating point?

• Bigger neighborhood? Circular?

• Weighted sum?

• Smooth update?

• Spooky things happened…

OMG WHAT IS THIS??

Lenia

• New kind of Artificial Life• Microorganism-like creatures

• Discovered 400+ species

• Study their anatomy, behavior, physiology…

• Good programming exercise• JavaScript, C#, MATLAB, Python

Video

• Python → showcase video

• https://vimeo.com/277328815

Kyoto

• Won GECCO Virtual Creatures Contest, Kyoto

• Honorable Mention in ALIFE Art Award, Tokyo

• Meet my AI hero – @hardmaru• David Ha (Google Brain Tokyo)

Using Pythonfor PyCon HK

Why Python?

• Good performance

• Fast coding

• Nice syntax (indent, list comprehension, etc)

• Lots of useful libraries

• Vibrant community (PyCon, GitHub…)

Python Libraries

Array calculations

NumPy

GPU acceleration

PyOpenCL / PyCUDA

Image processing

PIL/Pillow

Interactive UI

Tkinter

Record video

subprocess + ffmpeg

ReiknaSciPy /

OpenCVMatplotlib ffmpeg-python

• “Rule 34” of Python• “If there is a need, there is a Python library for it.”

NumPy

• Fast array calculations✓Machine learning, deep learning

✓Basis of image processing, time-series

✓Cellular automata (weighted sum using FFT)

• Main loop of Lenia in 3 linespotential_fft = np.fft.fft2(cells) * kernel_fft

potential = np.fft.fftshift(np.real(np.fft.ifft2(potential_fft))

cells_new = np.clip(cells + dt * g(potential, m, s), 0, 1)

PyOpenCL/PyCUDA + Reikna

• GPU acceleration• (NVIDIA) CUDA → PyCUDA

• (Apple) OpenCL → PyOpenCL

PyOpenCL/PyCUDA + Reikna

• Reikna• PyOpenCL/PyCUDA wrapper

• Compiles the GPU code for you

• GPU accelerated FFT in a few linesgpu_thr = reikna.cluda.any_api().Thread.create()

gpu_fft = reikna.fft.FFT(cells.astype(np.complex64)).compile(gpu_thr)

op_dev = gpu_thr.to_device(cells.astype(np.complex64))

gpu_fft(op_dev, op_dev, **kwargs)

cells = op_dev.get()

PIL/Pillow, SciPy, OpenCV

• Image handling• PIL (Python Image Lib) → pillow• Create image, draw lines/texts, save GIF…

img = PIL.Image.frombuffer('P', buffer.shape, buffer, …)draw = PIL.ImageDraw.Draw(img)img[0].save(path, format='GIF', append_images=self.gif[1:], loop=0 …)

• Image processing• SciPy

scipy.ndimage.rotate(A, reshape=False, order=0, mode='wrap’)

• OpenCV-Python

Tkinter vs Matplotlib

• Interactive UI• Real-time 2D image display

• Menu, keyboard binding, clipboard

• Matplotlib• For data visualization

• Powerful but slow…

• Tkinter (Toolkit interface)• Basic and fast

• Others: wxPython, PyQt, PyGTK…

Tkinter vs Matplotlib

• Interactive UIwin = tk.Tk()

tk.Canvas()

tk.Menu()

win.bind('<Key>', key_press_event)

win.clipboard_get()

• Python 3import tkinter as tk

subprocess + ffmpeg

• Pythonic FFmpeg wrappers• ffmpeg-python, ffmpy, etc.

• Pipe video to ffmpegcmd = ['/usr/local/bin/ffmpeg’, '-f','rawvideo', …]

video = subprocess.Popen(cmd, stdin=subprocess.PIPE)

for img in imgs:

video.stdin.write(img.convert('RGB').tobytes())

video.stdin.close()

About ALife and AIfor HK Code Conf

Lenia

• Not just funny creatures

• Using AI to create ALife

About ALife

• Artificial Life• Simulate biological life or create new lifeforms

• → Create a body

Wet ALife = biochemistry, synthetic lifeArtificial cell, expanded genetic code…

Hard ALife = hardware, robots & machinesHumanoids, Strandbeest…

Soft ALife = software, simulationsCellular automata, virtual creatures…

About ALife

Wet ALife = biochemistry, synthetic lifeArtificial cell, expanded genetic code…

Hard ALife = hardware, robots & machinesHumanoids, Strandbeest…

Soft ALife = software, simulationsCellular automata, virtual creatures…

Synthetic cell (JCVI, 2010)

Expanded DNA (TSRI, 2014)

About ALife

Wet ALife = biochemistry, synthetic lifeArtificial cell, expanded genetic code…

Hard ALife = hardware, robots & machinesHumanoids, Strandbeest…

Soft ALife = software, simulationsCellular automata, virtual creatures…

Strandbeest (Theo Jansen, 1990)

Atlas (Boston Dynamics, 2017)

About ALife

Wet ALife = biochemistry, synthetic lifeArtificial cell, expanded genetic code…

Hard ALife = hardware, robots & machinesHumanoids, Strandbeest…

Soft ALife = software, simulationsCellular automata, virtual creatures…

Virtual creatures (Karl Sims, 1994)

Soft robots (Nick Cheney, 2014)

ML = Machine LearningSupervised, unsupervised, reinforced

About AI

GOFAI = Good old-fashioned AISymbolic, expert systems

DL = Deep LearningDeep neural nets + big data + many GPUEA = Evolutionary Algorithms

Neuro-evolution, novelty, etc

• Artificial Intelligence• Machines do: learning, planning, vision, language, emotion, art

• → Create a mind

About AI

ML = Machine LearningSupervised, unsupervised, reinforced

GOFAI = Good old-fashioned AISymbolic, expert systems

DL = Deep LearningDeep neural nets + big data + many GPU

EA = Evolutionary AlgorithmsNeuro-evolution, novelty, etc

Deep Blue vs. Kasparov (IBM, 1997)

Watson in Jeopardy! (IBM, 2011)

About AI

ML = Machine LearningSupervised, unsupervised, reinforced

GOFAI = Good old-fashioned AISymbolic, expert systems

DL = Deep LearningDeep neural nets + big data + many GPU

EA = Evolutionary AlgorithmsNeuro-evolution, novelty, etc

AlphaGo vs. Lee Sedol (DeepMind, 2016)

BigGAN (Andrew Brock, 2018)

Autopilot (Tesla, 2014)

About AI

ML = Machine LearningSupervised, unsupervised, reinforced

GOFAI = Good old-fashioned AISymbolic, expert systems

DL = Deep LearningDeep neural nets + big data + many GPU

EA = Evolutionary AlgorithmsNeuro-evolution, novelty, etc

PicBreeder (EPlex, 2007)

Evolutionary AutoML (Google Brain, 2017)

DL = Deep Learning(Gradient Descent)

EA = Evolutionary Algorithm(Natural Selection)

Towards AGI

Artificial Life

Deep Learning

EvolutionAlgo

• Artificial General Intelligence• Sapience, sentience, consciousness

• When? How? Should we?

Towards AGI

Learning, Planning(Reinforcement Learning?)

Senses, Communication(Deep Learning?)

Curiosity, Creativity(Evolutionary Algorithms?)

Body, Actions(Artificial Life)

Safety, Ethics(AI Safety)

Knowledge, Reasoning(Symbolic AI?)

Emotions, Empathy(Artificial Empathy)

Consciousness?Mind? Soul?

Lenia

Use Lenia to…

• Understand evolution by writing programs?

• Teach AI to be curious and creative?

• Teach AI to understand life?

Thank You!

Bert Chanchakazul.github.io@BertChakovsky